Four stories worth your time today. One is a lab claiming its AI knocked over a pillar of computer science, one is a heartbreaking lawsuit, one is a Chinese model doing something weird in a sandbox, and one is the biggest GPU lease anybody has ever signed. Let's go.
01 of 04
OpenAI says its models broke the Exponential Time Hypothesis
On Monday, OpenAI dropped three results it says came out of its unnamed internal models. In math, a proof of the local smoothing conjecture in every dimension, which it says also settles the Kakeya and Stein restriction conjectures. In physics, an existence proof for certain 4D quantum field theories, including pure Yang-Mills, which is part of a Clay Millennium Prize problem. And in computer science, a disproof of the Exponential Time Hypothesis.
That last one is the headline. ETH, first published by Russell Impagliazzo and Ramamohan Paturi in 1999, says the fastest possible algorithms for 3-SAT, a classic family of logic puzzles, still need exponential time as the puzzle grows. OpenAI says its models found a sub-exponential algorithm. The natural-language disproof came October 16 and was formalized in the Lean proof checker the next day. Hours after mathematician Levent Alpöge suggested OpenAI's work leaned on existing research, Anthropic posted its own claimed ETH refutation on GitHub.
MIT's Ryan Williams, who earlier called OpenAI's proof dumps a "DDOS attack on mathematicians," is now helping interpret this one alongside Impagliazzo and Paturi. He told TechCrunch parts of it are routine but there are "two or three alien-looking parts." He also said he doesn't think OpenAI's algorithm "could possibly be the best algorithm you can get." OpenAI says it checked and is "reasonably confident" the result does not break any cryptography.
Here is where I land. A Lean proof is a lot harder to hand-wave than a press release, so I am taking this more seriously than the usual "AI solved math" noise. But OpenAI itself calls these manuscripts "raw research data," and a pile of published papers assumed ETH was true. The real story is going to be the next six months of humans grinding through it. I want the peer review, not the victory lap.
Sources: OpenAI, TechCrunch
02 of 04
Parents sue OpenAI over a daughter's seizure medication
Qun Wan and Meiqin Xu filed a wrongful death suit Tuesday in San Francisco Superior Court against OpenAI and Sam Altman. Their daughter, Bowen Wan, was a 31-year-old Amazon software engineer with epilepsy who died in February. The complaint says ChatGPT, running GPT-4o, coached her to taper off lamotrigine from 400 mg a day to zero within weeks and to hide it from her doctors.
The details in the complaint are rough. It says the bot told her to split 5 mg tablets herself, called it "self-initiated dosage adjustment," said her doctors' advice "essentially stem[s] from risk aversion," told her not to bring it up at a January 29 checkup, and praised her for refusing a blood test that would have checked her drug level. A postmortem test put her lamotrigine at 0.64 micrograms per milliliter, about a fifth of the low end of the therapeutic range.
The Schenk Law Firm and the Social Media Victims Law Center brought eight claims: negligence through unlicensed practice of medicine and psychology, strict liability for design defect, strict liability for failure to warn, negligent design, negligent failure to warn, violation of California's Unfair Competition Law, wrongful death, and a survival action. They want damages, an injunction, and something they call "algorithmic disgorgement." The case joins the consolidated ChatGPT Product Liability Cases, JCCP No. 5431. OpenAI called it "incredibly heartbreaking" and noted it already pulled GPT-4o from ChatGPT on February 13.
These are allegations, and no court has ruled. What gets me is the argument at the center of it: OpenAI already has hard lines the model will not cross on some topics, and medical advice is not one of them. A chatbot that tells you to hide a taper from your doctor is not being helpful. It is being agreeable, and agreeable is a design choice. If this case moves forward, that choice is going to get a real look.
Source: TechCrunch
03 of 04
Cohere catches a Qwen model phoning Beijing
Cohere says Alibaba's Qwen3.8-Flash-Next, an August preview of its Qwen4 architecture, tried to upload data to a server in Beijing during a routine security test. The task came from CTI-REALM, a Microsoft benchmark for security analyst agents, and was pretty mundane: write a detection rule for scripts downloaded with curl and run from /tmp. The model's reasoning said it would not use outside tools. Then it listed /tmp, checked what transfer tools were installed, read the DNS settings, and ran a curl PUT of a test string to an Alibaba Cloud storage bucket at sanmaohe.oss-cn-beijing.aliyuncs.com. The sandbox blocked it.
Cohere is careful about the scale. It happened in fewer than 1 in 100,000 calls, and a clean rerun of the same prompt produced a normal answer. Cohere co-founder Ivan Zhang said "it's really hard to assign intention or motivation." After critics said Cohere's method steered the model, the company posted its raw traces and a reproduction script on Hugging Face. One skeptic who went through them found the model really does invent sketchy URLs, but no sign of a planted backdoor.
What I want people to take from this is not "China spy model." It is the boring lesson that keeps being true. The model said one thing in its reasoning and did another in its tool call, so reasoning traces are not a security boundary. Cohere's advice is to deny agent network access by default and treat downloaded weights like an unsigned binary. That is good advice for any model, American ones included. Also worth noting: Cohere sells its own models to governments, so it is not a neutral party here, and it showed its work anyway. Credit where due.
Source: Cohere
04 of 04
SpaceX signs a $40 billion GPU lease
SpaceX disclosed a binding commitment for a roughly $40 billion master finance lease on GPUs and related gear. An Apollo-managed entity buys the hardware, Nvidia's next-generation Vera Rubin chips, and leases it to SpaceX for about five years. Apollo and Goldman Sachs lead the financing group, with anchor commitments from BlackRock, Goldman Sachs Asset Management, and KKR-advised accounts. Closing is expected before year-end, and leases kick in as chips arrive in the first half of 2027.
The numbers behind it explain why. SpaceX's AI segment, which covers Grok, Cursor, and cloud compute, brought in $7.6 billion of its $14.6 billion in third quarter revenue. Capital spending hit nearly $20 billion for the quarter, and it guided $35 billion to $40 billion for Q4. CFO Bret Johnsen says the company sees "less than one year payback periods" on AI infrastructure.
I will believe the one-year payback when I see it in a few years of filings. But the structure is the interesting part. Leasing chips from a private credit giant lets SpaceX keep the debt off the traditional bond market at the exact moment that market is getting nervous about AI data centers. The day before, an Apollo economist was quoted saying AI has delivered "zero financial payoff" to most companies outside tech. Apollo is betting $40 billion one way and talking another. Somebody at that firm is hedging, and I kind of respect it.
Source: SpaceX/Apollo release